* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download GENES in the Optimization
Genome evolution wikipedia , lookup
Biology and consumer behaviour wikipedia , lookup
Pharmacogenomics wikipedia , lookup
Koinophilia wikipedia , lookup
Quantitative trait locus wikipedia , lookup
Medical genetics wikipedia , lookup
Designer baby wikipedia , lookup
Genetic drift wikipedia , lookup
Behavioural genetics wikipedia , lookup
Heritability of IQ wikipedia , lookup
History of genetic engineering wikipedia , lookup
Genetic code wikipedia , lookup
Gene expression programming wikipedia , lookup
Human genetic variation wikipedia , lookup
Genetic engineering wikipedia , lookup
Public health genomics wikipedia , lookup
Population genetics wikipedia , lookup
Genetic testing wikipedia , lookup
Microevolution wikipedia , lookup
Beam Dynamics study by using Genetic Algorithms and the ELI-np case Alberto Bacci – INFN-Milano Alberto Bacci, Laboratoire de L’Accélérateur Linéaire (Orsay – France), 13th June Outline Genetic Algorithms (GAs) introduction Historical notes Hints on What are & Why use GAs Where: Genetic Algorithms in Beam Dynamics Optimization GAs applied to Beam-Lines Optimization From Beam-Lines to Chromosomes Following Genetic Laws: Fitness, Reproduction, Mutations, … E.G.: SPARC beam line Optimization in Thomson case The GIOTTO code The Data-Based DB Inputs & Outputs Fitness function (or Idoneity) definition Optimizations & Statistics on Specific Cases: Ultra short bunches by Hybrid Velocity Bunching Comb bunches distributions Eli-np Case: The reference Working Point & Statistic Genetic Algorithms (GAs) introduction: Historical Notes 1970 John Holland - schemata theorem John Holland 1975 J. Holland publication: “Adaptation in Natural and Artificial Systems: An Introductory Analysis with Application to Biology, Control, and Artificial Intelligence”. The Seminal work 1975 K. De Jong (J. Holland’s student), Thesis: “An analysis of the behavior of a class of genetic adaptive systems”. Broad applicability of GAs 1989 David Goldberg Book: “Genetic Algorithms in Search, Optimization, and Machine Learning” It deals with the topic at high level and is considered a milestone in GAs story. It reports techniques like Multi Objective GA (MOGA), today very current. Introduction: What & Why What are Genetic Algorithms (GAs) Searching procedures based on natural selections (genetic laws) Why choosing GAs versus other techniques … A basic answer: Newton-Raphson methods (and many variants) are based on local information. The Scan moves in direction of “local maxima” or “local minima” Local extremaA Parabola start start End End Introduction: What & Why Why Genetic Algorithms ** Despite, Newton-Raphson methods can overcame the local solutions issue by some tricks … GAs are: naturally able to manage the local solution issue naturally parallelizable usable with a minimum mathematical effort And, by empiric results, show strong capability to manage problems where other methods fail ** pros and cons in literature Where: Genetic Algorithms (GAs) in Beam Dynamics Optimization -1 GAs give strong advantages … in multi-dimensional problems with variables strongly non linear correlated A main example : Space Charge & its non-linear nature: correlates low energy beam-line parameters Also frozen beams (space charge off): Other complex situations Example (1) Thomson/Compton Sources (e.g. SPARC_lab, STAR, ThomX, ELI-np, Munich Compact Light Source) which ask for : For the spectral density: 1) very low DE/E, 2) low Emit For the photon flux : 3) Qbunch as high as possible Where: … - 2 Example (2) Ultra short e-beams (e.g. SPARC_lab, LCLS, REAGE, XFEL, EUPRAXIA, … ): Femtosecond light pulses (FEL/X-FEL), Atoms in chemical reactions, phase-transition, Photosynthesis Water Splitting : timescale 1-100 fs [2014 “first snapshots of water splitting” by LCLS; ScienceDaily; Nature] Plasma Wave Acceleration: λplasma order of 30-600 fs . The Witness much shorter, the Driver (pwfa) comparable to λplasma Femtosecond Electron Diffraction (FED) Molecular or atomic motion movies: phase transitions, …, . Timescale: few 10s of fs. Relativistic case: Eb ≈ 5MeV, Qb ≈ 100fC, εtr<0.1 mm-mrad, σz<30μm (100fs) THz radiation (by CoTrRad) 0.1 up to tens THz is of great interest for both longitudinal electron beam diagnostics (fs scale) and spectroscopy in pump-and-probe experiments .... Genetic Algorithms applied to Beam-Line Optimization From Beam-Lines to Chromosomes Genetic Laws work on Chromosomes ==> Chromosomes are made of genes (parameters) Beam-Line = Parameters Array ==> One Chromosome Chromosome A Beam-Line = Following Genetic Laws: Fitness Function Rules to pass generation in generation: Selection: bluffed Roulette wheel Mutations …. & others methods & tricks. The rule: closest to Nature, best performances Chromosomes Must be sorted by a Fitness function == 50 e εn,x eimtX 0.35 + 2 50 e E spread 250 Δγ/γ 2 Following Genetic Laws: Reproduction & Mutation Chromosome Sorting by 0< < 100 By the Roulette Wheel: two Chromosomes Tracking code sum of probability == 1 99. 97. 90. 85. 60. 58. Chromosomes reproduction Gun gradient Gun Ф injection Bz intensity first solenoid Mutation, with probability < 1% Following Genetic Laws: Real coding (1) and binary coding (2) Chromosome Main loop, Generation in Generation (1) Genetic rules on Real Numbers 0 < Gene Value < 255.127 0|0|0|0|0|0|0|0. 0|0|0|0|0|0|0 (2)Binary Coding 1|1|1|1|1|1|1|1 . 1|1|1|1|1|1|1 8 bits As DNA in Genetic Laws Genetic rules on binary arrays 7 bits 15 bits X Gene Decoding to decimal and compute the fitness Main loop, Generation in Generation Following Genetic Laws: Elitism > Concluding: the elitism Sorting oper Nowadays “quasi-classic” optimization techniques o elitism o advanced mutation operators o hill climbing o regeneration from best solutions o ……… o & parallelization quasi-mandatory +2 +5 E.G.: SPARC beam line Optimization in Thomson case Old Kind of Fitness Function (a) By hand (b) By GA F fitness 1 x ,n spot “MAXIMIZING THE BRIGHTNESS OF AB ELECTRON BEAM BY MEANS OF A GENETIC ALGORITHM” A. Bacci, C. Maroli, V. Petrillo, A. Rossi, L. Serafini - NIMB 263 (2007) 488-496 spectrum The GIOTTO code GIOTTO - Genetic Interface for OpTimising Tracking with Optics WAS BORN in 2008; Language: Fortran 90/95 USE for Optimization of Generic Code’s Parameters or for Statistical (Jitters) Analysis INPUTS based on NameList & on two internal DataBase CAN easily Drive different codes: Now: ASTRA’s Generator, ASTRA, QFluid (Plasma acceleration, A. Rossi modifications) Current Version (Ver. 10.0): Linux 64 bit; Windows 64 – (compilers gfortran or INTEL fortran) Parallelized with OPEN-MPI (Linux), MS-MPI or INTEL MPI (Windows) Used @ PSI (S. Bettoni) and tested at Desy-Pitz Code and Documentations: URL: http://pcfasci.fisica.unimi.it/Pagine/GIOTTO/GIOTTO.htm (server down, pardon!) Exist an User manual for version 8.5 2012 (needs updates) GIOTTO – Genetic Interface for OpTimising Tracking with Optics From 2008 up to day, the code is grew in power and capability Optimization techniques: elitism; advanced mutation operators; hill climbing; ant colony regeneration from best solutions user freely defined by Astra outputs: Targets: bunch PosZ or Time, En, Enspread, SigZ, Xemit, sigX, divergX, Yemit, …. Important GIOTTO’s features: Every NameList ’s variables can be used as a GENE (optimizable) & Any code working with NameList is directly importable in GIOTTO. ASTRA e.g. : Phi(1)…Phi(50), MaxE(1:50), MaxB(1:50), sig_x (laser cathode) ,sig_clock (Laser @ cathode), …, … (no limit on the number) Constraints freely defined by the user (under test) Switch from Optimizations to Statistical analysis is really EASY Jitters sampling interval: Uniform or Gaussian GIOTTO’s Data-Bases All variables usable in THE BEAM-LINE Now: Astra_generator, Astra, QFluid DB_1 Optimizable variables THE GENEs Sub-DB_1 Astra Outputs (or other codes) THE FITNESS emittance, envelope, En_spread, etc. … DB_2 GIOTTO’s INPUT FILE GINxx.xx.ini is divided in two parts: 1) One NameList (&GA) giving all the directive to GIOTTO 2) 1) 2) Three keyNames defining: CONSTRAINTS, FITNESS and GENES GIOTTO’s INPUT FILE: &GA NameList Under developing (it slows down heavily GIOTTO) Usually few variables are used Optimization Rarely needs changes GIOTTO INPUT FILE: Key_Names Rarely needed Definition Comments GENES Definition Comments GIOTTO: FITNESS FUNCTION Reverse Polish Notation: Opers Follow Operands 3 4 + = 7 Stack based operation Does not need brackets = 50 ∙ 𝑒 − 𝑒𝑚𝑖𝑡𝑋 2 2.5 +50 ∙ 𝑒 − 𝑠𝑖𝑔𝑍 2 0.15 +50 ∙ 𝑒 − 𝑠𝑖𝑔𝑋 2 2.5 a) not yet full optimized b) full optimized Fitness Funciont strategy to Cope with Multi Objectives Problems (MO): o One Single Criterium per Equation piece (Objectives Wights) o To be close to the Goal mean close to the Gaussian Curve Top o The ‘Far region’ (referring to optimization) has to be on the maximum Gaussin slope o change the FF in real time (ander implementation) GIOTTO Optimization & Statistical analysis Beam-Line Optimization for: ultra short, ultra cold, High brightness bunches 2.0 m drift C-band cavity GIOTTO RESULT Drift @ 20 MeV Envelop[mm] Sig_z[um] ΔE =100 keV Emit[um] GOALS: Low Emittance Low Energy Spread femptosec. Sig_Z A Beam-Line studied with: Experience An Ad hoc GIOTTO use GENES in the Optimization: Gun: • (1) Phase & (2) Solenoid (Bz) TW cavity (RF- Compressor): • (3) Phase & (4) Solenoids C-band cavity: (5) Phase GENES GIOTTO OUTPUT: RISULTATI.TXT RISULTATI.TXT Emit Id value Best Chromosome of the Generation N.5 Generation sigZ ∆𝑬 Beam-Line STATISTIC for Laser Comb (ECHO Bunch Generations) 6 Bunches 4 Bunches - P.O.Shea et al., Proc. of 2001 IEEE PAC, Chicago, USA (2001) p.704. - M. Ferrario. M. Boscolo et al., Int. J. of Mod. Phys. B, 2006 (Taipei 05 Workshop) 360 cases Beam-Line STATISTIC for Laser Comb – TWO Bunches case : Current Statistic GIOTTO and the ELI-NP case ELI-NP LINAC machine Beam Dynamic features Electron Linac design to drive bright Compton back-scattering gamma-ray sources A. Bacci, D. Alesini, P. Antici, M. Bellaveglia, R. Boni et al. J. Appl. Phys. 113, 194508 (2013); online: http://dx.doi.org/10.1063/1.4805071 The Peculiar Gamma Ray Source features In next page, how to reach these parameters ELI-NP Injector merit factors The Emittance Maximization of electron density into transverse phase space : ==> means ==> very low emittance ̴ 0.4 mm-mrad The Energy Spread Minimization of the energy spread: the source spectral density require Δγ/γ < 0.1%, we have chosen a conservative threshold of 0.05 % Energy Spread by RF curvature: σz < 280 µm @ the injector exit Booster freq. Result: GIOTTO optimization on merit factors ( booster’s injector) Comparison using Tstep (a Parmela heir) & Astra codes 1) A space charged dominated region needs a double check 2) Astra gives the possibility to use Giotto (Giotto improves 30-60 %) Astra with GIOTTO Tstep Astra Tstep γε =0.4 μm <E>=79.8 MeV σz= 0.279 mm σE/<E> = 1.65% 31 Injector jitters analysis a full S2E: Injector→ C Booster → γ-Source Astra→ Elegant → Cain ELI-NP booster’s Injector Jitters (9 parameters) Jitters kept in Consideration in GIOTTO: RF: 200 fs Phase (overestimated): (1) Gun & (2,3) TW cavities (S- band) 2‰ in pick field: (4) Gun & (5,6)TW cavities (S- band) laser: (7) 200 fs arrival time (8) 20 µm pointing instabilities (on cathode) (9) 5% energy fluctuation (Charge fluctuation) Different Machines: +/- 70 µm as misalignments for: RF Cavities, Gun Solenoid, TW Solenoid Uniform distributions Jitters; a very conservative choice ELI-NP Injector Jitters Analysis – Energy & Bunch length All jitters σz Machine_1 160 runs Machine_2 160 runs Machine_3 160 runs 34 ELI-NP Injector Jitters Analysis – Centroid, Envelope, Emittance All jitters Machine_1 160 runs Machine_2 160 runs Machine_3 160 runs 35 In Conclusion Genetic Algorithms show great promise in the Beam Dynamics optimization and problem solution. GIOTO has been applied successfully to: • refine known beam lines, with improvements around 20-40 % (in the performances) • have been used to find completely new schemes, as in case of the hybrid velocity bunching Thanks for your attention Demand of EXTREME HIGHT QUALITY electron beams doesn’t stop and often it makes necessary to cope with strong space charge Beam-Line optimization is Nowadays really a critical issue Emittanzometro RF-Gun 1.6 S-band 120 MV/m Emittanzometro’s data are handled by a dedicate algorithm that return an intensity matrix PhaseSpace.txt (successively interpolated to increase the definition) Emittance in Real Beam Since real beams usually do not have well defined boundaries, a method for calculating the emittance, is to choose a specific density contour, in the phase space, that represents from the 50% (worst cases) up to the 98-99% (best cases) of the whole bunch charge (or integrated intensity). Within this density contour and under certain conditions, such emittance satisfies the Liouville’s theorem and thus is conserved 100% 97% 95% x 2 y 2 2xy gx 2 by 2 2axy 1 2 1 gb a 2 area GMESA normalization g 1 2 , b , a gx 2 bx2 2axx 1 gx 2 bx 2 2axx 1 gx 2 by 2 2axy 1 g , b , a One of the 30 generation’s chromosome with 25 genes c g b a g b a 30 1 2 g b a 8 The space phase ellipse’s sampling is a thorny issue as uniformity and dimension. A shuffled uniform random generator is used. I F fitness Ad Data analysis at SPARC - Some relevant results One of the more significant curve - analyzed also with two other methods - that shows a strongly marked double minimum Some phase space of the emittance curve From first tests the code seemed to be able to analyze the rough phase space images, not yet interpolated An output file 4 BPM and Genetic control of the steering correctors A code to compute the rms emittance of high brightness electron beams Developing of a dynamic link library (dll), in fortran 90, that can interface with LabView Current of maximum variation per corrector N. Corrector rms of the BPM off-set Sample of the Correctors configurations Generation n best solution Generation n+1 memory of the best solution from simulations it converges to 0.200 mm of maximum off-set after 80 generation Considering 2s to test each configuration → 36 minutes The real world could be faster ant colony optimization